Fix for regression after Global Load Scalarization patch

[• tstellarAMD]

test/CodeGen/AMDGPU/mesa_regression.ll

4

There should be a CHECK line here to check for the expected load instruction.

[alex-t]

test/CodeGen/AMDGPU/mesa_regression.ll

4

No. This patch fixes compiler error - not incorrect code generation.
With this change test can be compiled w/o errors with -amdgpu-scalarize-global-loads=false. LIT reports test as PASSED. W/o change compilation aborts with error and LIT reports test as FAILED.
Presence of any vector or scalar loads in output does not test anything.

[• tstellarAMD]

test/CodeGen/AMDGPU/mesa_regression.ll

4

The original error was 'Cannot Select...', so we should add a check line that verifies we are able to select this node now. This is what we've always done when adding tests for these kinds of failures.

[alex-t]

Test changed

[kzhuravl]

lib/Target/AMDGPU/SIISelLowering.cpp

2788–2790

Can we check Subtarget->getScalarizeGlobalBehavior() first? Other checks are more expensive. Also formatting.

[alex-t]

lib/Target/AMDGPU/SIISelLowering.cpp

2788–2790

Yes. Good point about the check order. Thanks

[arsenm]

I think it would be better to just add run line to the original test with the option disabled

test/CodeGen/AMDGPU/mesa_regression.ll

1

You don't need -O2

[alex-t]

In D27750#622326, @arsenm wrote:

I think it would be better to just add run line to the original test with the option disabled

Exactly! ) This is the point I started with. The option is "false" by default and I considered that we don't need separate test for this.

[arsenm]

Looks like this was r289822

lib/Target/AMDGPU/SIISelLowering.cpp

case AMDGPUAS::CONSTANT_ADDRESS:
     if (isMemOpUniform(Load))
       return SDValue();
     // Non-uniform loads will be selected to MUBUF instructions, so they
     // have the same legalization requirements as global and private
     // loads.
     //
     LLVM_FALLTHROUGH;
   case AMDGPUAS::GLOBAL_ADDRESS: {
-    if (isMemOpUniform(Load) && isMemOpHasNoClobberedMemOperand(Load))
+    if (Subtarget->getScalarizeGlobalBehavior() && isMemOpUniform(Load) &&
+                  isMemOpHasNoClobberedMemOperand(Load))
       return SDValue();
     // Non-uniform loads will be selected to MUBUF instructions, so they
     // have the same legalization requirements as global and private
     // loads.
     //
   }
     LLVM_FALLTHROUGH;
   case AMDGPUAS::FLAT_ADDRESS:
     if (NumElements > 4)

test/CodeGen/AMDGPU/mesa_regression.ll

+; RUN: llc -O2 -mtriple amdgcn--amdhsa -mcpu=fiji -amdgpu-scalarize-global-loads=false -verify-machineinstrs  < %s | FileCheck %s
+
+; CHECK-LABEL: %entry
+; CHECK: flat_load_dwordx4
+
+define amdgpu_kernel void @store_global(<16 x double> addrspace(1)* nocapture %out, <16 x double> addrspace(1)* nocapture readonly %in) {
+entry:
+  %tmp = load <16 x double>, <16 x double> addrspace(1)* %in
+  store <16 x double> %tmp, <16 x double> addrspace(1)* %out
+  ret void
+}